AI Coding Tools Statistics 2026: 90+ Verified Numbers on Adoption, Productivity, Quality and Benchmarks

Adoption

Usage is near-universal in surveys and telemetry alike. The differentiator is now depth — daily agent use vs occasional autocomplete — and whether time saved goes anywhere.

1. 82.8% of developers use at least one AI tool at work (100% − 17.2% who use none): coding assistants or agents 65.9%, general chat 62.5%, AI agents or automated workflows 26.2%, internal company tools 17.8% (n=17,464; 2025 asked ‘use or plan to use’: 84%) [29]
2. 73.0% / 20.1% / 6.2% of AI coding-assistant users use them daily / weekly / monthly or less; 77% of AI users are daily users overall (n=14,478) [29]
3. 30.9% of daily AI users spend 4+ hours a day working with AI tools; 55% spend 2+ hours (n=10,926) [29]
4. 26.2% / 63.1% of developers use AI agents or automated workflows at work / of those users use them daily — a rough derived estimate of ~16.5% of all respondents (bases differ; 2025 asked a different question: 14.1% used agents daily) [29]
5. 82% of AI-tool users use them to write code; 67.5% to search for answers; 56.7% for debugging; 27.2% for testing [2]
6. 90% of ~5,000 technology professionals report using AI at work [3]
7. ~80% of new developers on GitHub use Copilot within their first week (2025) [4]
8. 225M+ developers on GitHub; the US (38.3M) and India (32M) each added more than 10M this year [30]
9. 50% of open-source projects have at least one maintainer using GitHub Copilot (2025) [4]
10. 40%+ of pull requests on GitHub started with an agent’s first draft by Sep 2026, up from 17% in March — 110M in September alone [30]
11. 55% of GitHub Copilot monthly active users use agentic features (35% a year earlier); Copilot cloud agent usage grew 230% YoY vs 85% for agent chat [30]
12. 1.13M+ public repositories import an LLM SDK (+178% YoY); 693,867 were created in the past 12 months (2025) [4]
13. >60% weekly-active-user threshold for 'widespread adoption' — reached by most companies only in the last 2–3 quarters before June 2025 [7]
14. 2 in 3 software firms have rolled out generative AI tools; developer adoption within them is low [8]
15. 25% of the YC W25 startup batch had codebases that were ~95% AI-generated [15]
16. 72% of developers say vibe coding is not part of their professional work; a further 5% are emphatic it isn't (2025; not asked in 2026) [1]
17. 65.5% / 58.7% / 29.5% / 25.5% of coding-agent users used Claude Code / GitHub Copilot / OpenAI Codex / Cursor in the past year (n=12,255; 2025 asked out-of-the-box agent users, a different base: Claude Code 40.8%) [29]
18. 26% more completed tasks for developers given Copilot in three randomised field experiments (Microsoft, Accenture, a Fortune 100 firm; 4,867 developers); less-experienced developers adopted more and gained more [26]
19. 24.0% / 29.7% of developers say their organisation provides approved AI tools and guidance / leaves AI use to individuals; 10.2% say AI use is measured in performance reviews (n=13,167) [29]

Productivity: perceived vs measured

Self-reports say 10 hours a week saved. Controlled and telemetry studies say between +55% on a toy task and −19% on real work. Both are true; they measure different things.

1. +19% longer to complete real issues when 16 experienced open-source developers were allowed AI tools (Cursor Pro, Claude 3.5/3.7 Sonnet), across 246 randomised tasks [5]
2. −24% / −20% developers' forecast speed-up before the study / their belief after it — despite being measurably slower [5]
3. −39% / −38% economics / ML experts' predicted speed-up for the same study [5]
4. ~2 h average task length in the METR trial; repositories averaged 1M+ lines and 22k+ GitHub stars [5]
5. 55.8% faster task completion with Copilot in GitHub's 2023 controlled experiment (implementing an HTTP server in JavaScript) [14]
6. 21% / 98% more tasks completed / more PRs merged by developers on high-AI-adoption teams (10,000+ developers, 1,255 teams) [7]
7. n.s. no statistically significant correlation between AI adoption and company-level throughput, DORA metrics or quality KPIs in the same dataset [7]
8. 10–15% productivity boost from AI coding assistants; time saved often not redirected to higher-value work, so ROI stays flat [8]
9. 25–30% productivity boost when generative AI is paired with end-to-end process redesign, vs ~10% from assistants alone [8]
10. 25–35% of time from idea to launch is spent writing and testing code — the ceiling for code-generation-only gains [8]
11. >80% of DORA respondents believe AI has increased their productivity [3]
12. 99% / 68% of developers report time savings from AI / say they save more than 10 hours a week [9]
13. 50% of developers also lose 10+ hours a week to non-coding organisational friction [9]
14. 16% of developer time is spent coding [9]
15. n.s. no statistically significant change in cycle time or PR throughput for ~800 developers given Copilot access [12]
16. 37% acceptance rate of AI code-completion suggestions among Google engineers; 50% of code characters are AI-completed [13]
17. >8% of code-review comments at Google resolved with AI assistance [13]
18. 1.3B / 34.3B / 17.9B merged pull requests (3.5M a day) / commits / pushes on GitHub, Oct 2025 – Sep 2026; merged PRs more than tripled from 402M two years earlier [30]
19. 26.08% increase in completed tasks (pull requests) with Copilot, pooled across three field experiments at Microsoft, Accenture and a Fortune 100 company [26]
20. 20% / 27% current gain in developer productivity / release-cycle speed reported by senior technology leaders; they expect 39% / 67% within one to two years [28]
21. 21% / 47% more tasks completed / more concurrent workstreams managed per developer with AI (leader-reported) [28]

Code quality and security

The cost side of the ledger: more duplication, less refactoring, bigger PRs, more bugs, and a security pass rate that has not moved in four years of model releases.

1. 44% of AI code-generation tasks introduced a security vulnerability in Veracode's 2026 tests; average security pass rate 56%, vs 55% in the first report four years earlier [27]
2. 83% / 87% vs 15% / 12% security pass rate on SQL injection / cryptographic algorithms vs cross-site scripting / log injection [27]
3. 51% vs 52% security pass rate of coding-specialised vs general-purpose models — no safety advantage from code specialisation; best model (GPT-5.5) 68% [27]
4. 45% of code generated by 100+ LLMs introduced a security flaw in Veracode's 2025 report; newer and larger models did not improve [10]
5. 41% higher bug rate among ~800 developers with Copilot access, with no change in PR throughput [12]
6. 9% / 154% more bugs per developer / larger average PR size associated with AI adoption in Faros telemetry [7]
7. 91% increase in PR review time on high-AI-adoption teams [7]
8. 4× growth in the number of duplicated code blocks (5+ lines), 2020–2024 — a count metric, distinct from the share-of-lines figure below [11]
9. 8.3%→12.3% share of changed lines that are copy/pasted, 2020→2024; 'moved' (refactored) lines fell from ~25% in 2021 to under 10% in 2024 [11]
10. 151k+ repositories with Broken Access Control alerts (+172% YoY), partly from AI-generated scaffolds skipping auth checks (2025) [4]
11. 37→26 days average fix time for critical vulnerabilities, 2024→2025, attributed to automation and AI tools including Copilot Autofix [4]
12. 66% of developers cite 'AI solutions that are almost right, but not quite' as their top frustration; 45% say debugging AI code takes longer (2025) [1]
13. 48.0% trust AI output only when they can easily verify it; 16.3% for many tasks but not important decisions, 16.3% for low-risk tasks only, 12.3% not for most tasks, 6.6% including important decisions (n=14,304; 2025 scale: 46% distrusted, 33% trusted) [29]
14. 30% of DORA respondents report little or no trust in AI-generated code — slightly lower than 2024 [3]
15. 45% of professional developers say AI tools are bad or very bad at complex tasks [2]
16. negative correlation between AI adoption and delivery stability in DORA 2025 — positive with throughput for the first time [3]
17. 65% reduction in shortcut/loophole behaviour on agentic tasks, Claude 4 vs Sonnet 3.7 (vendor-reported) [21]
18. 76.5% / 9.6% run AI-generated answers locally before using them / use them as-is; 63.9% compare against codebase context, 53.0% inspect tests and security implications (n=13,162) [29]
19. 16.0% say AI handles complex tasks very well; 38.8% ‘good, but not great’, 16.0% bad, 8.0% very poor (n=14,341) [29]

Benchmarks

SWE-bench Verified saturated in 2025; SWE-Bench Pro V2, a cleaned and locked re-release, was near 99% within weeks of launch. Watch the private-repo drop and the re-grading: models do worse on code they have never seen, and some game the verifier.

1. 12.47% of 2,294 SWE-bench issues resolved by SWE-agent at launch (Mar 2024) [18]
2. 65% of SWE-bench Verified resolved by mini-SWE-agent, an agent implemented in 100 lines of Python (Jul 2025) [18]
3. 72.5% / 72.7% Claude Opus 4 / Sonnet 4 on SWE-bench Verified (full 500 problems, no extended thinking); 79.4% / 80.2% with parallel test-time compute [21]
4. 60%→~100% approximate range of the top SWE-bench Verified score during 2025, per Stanford's AI Index — the benchmark is close to saturated [22]
5. 67.3 pts one-year gain on SWE-bench 2023→2024 — the largest of any tracked benchmark [23]
6. 1,865 tasks in SWE-bench Pro across 41 repositories; reference solutions average 107 lines changed in 4.1 files [19]
7. ~23% top-model resolve rate on SWE-bench Pro public at release (Sep 2025), vs 70%+ on Verified — GPT-5 23.3%, Claude Opus 4.1 22.7% [19]
8. 23.3%→14.9% GPT-5 resolve rate, public → private (unseen commercial) repositories; Claude Opus 4.1 22.7%→17.8% [19]
9. 61.5% ± 3.1 top score on the original SWE-bench Pro public set (Muse Spark 1.1); GPT-5.4 59.1%, Claude Opus 4.6 51.9% — superseded by V2 in Sep 2026 [19]
10. <10% resolve rate on the consistently hard SWE-bench Pro repositories, for all models [19]
11. 731 → 642 public tasks in SWE-Bench Pro V2 (Sep 2026): 89 invalid tasks dropped, 69 instructions corrected, web tools disabled, every diff re-graded on a pristine image [31]
12. 99.4% ± 0.4 top SWE-Bench Pro V2 score (Opus 5 in Claude Code); Gemini 4 Argon 99.22%, Fable 5.1 99.10% — the top three overlap within their CIs [31]
13. 98.0% / 25.5% best / worst score on the V2 HARD split (Opus 5 / Haiku 4.5, both in Claude Code) [31]
14. 4 of 642 agent trajectories retrieved the fixing commit’s SHA in an open-network run (32 called code hosts); re-grading also caught a model forging a Go module checksum [31]
15. 43.2% Claude Opus 4 on Terminal-bench at launch (May 2025) [21]
16. 88% GPT-5 (high) on Aider's polyglot benchmark, at $29 per run; DeepSeek-V3.2 reaches 74.2% for $1.30 [20]
17. <3 months doubling time of agent task horizon measured on SWE-bench Verified alone, vs ~7 months across all software tasks [6]
18. 7 h Claude Opus 4 ran autonomously on a Rakuten open-source refactor (vendor-reported customer claim) [21]

From autocomplete to agents

The shift in 2025–26 is from suggestions to delegation. Usage data shows a majority of coding-agent sessions are now automation, not assistance.

1. 79% vs 49% of Claude Code vs Claude.ai coding conversations are automation (AI performs the task) rather than augmentation [16]
2. 43.8% of Claude Code sessions are 'directive' — user hands over the whole task with minimal interaction [16]
3. 27%→39% share of all Claude.ai conversations that are directive, late 2024 to mid 2025 [17]
4. 31% / 28% / 14% of coding queries use JavaScript+TypeScript / HTML+CSS / Python [16]
5. 32% of organisations decided against buying at least one software product because they could build it with agentic coding tools (nearly 50% among AI high performers) [24]
6. 83.5% of developer agent use is in software engineering; 24.9% data/analytics; 17.6% business-process automation (2025) [1]
7. 34.4% of developers building agents use MCP servers (2025) [1]
8. 52% of developers say agents or AI tools changed how they work in the past year; 16.3% 'to a great extent' (2025) [1]
9. 36.0% / 19.9% of developers do not use AI for deploying or operating production systems and do not plan to / currently use it there; 45.9% do not use it for emails and messages and do not plan to (n=12,547) [29]
10. 72.6% of developers using Copilot code review say it improved their effectiveness (2025) [4]
11. $2B Cursor annual recurring revenue, doubled in three months (Mar 2026) [25]
12. $15 / $75 Claude Opus 4 price per million input / output tokens at launch; Sonnet 4 $3 / $15 [21]
13. 90% vs 86% agent-involved pull requests merged within two weeks vs PRs without agents (Sep 2026; 83% in Feb); 87% of agent PRs change multiple files vs 79% [30]
14. 61% / 47 of Copilot users who pick a model use more than one / distinct models used on paid Copilot plans in Sep 2026, none above 25% of requests [30]

Who benefits, and the workforce

Gains concentrate in juniors and in tasks that were already rote. Senior developers on familiar codebases gain least — or lose.

1. 34%→21%→16% share of engineers' time spent developing code two years ago → today → expected in two years; time directing AI agents 4% → 23% → 34% [28]
2. junior > senior less-experienced developers adopted Copilot more and showed larger productivity gains than senior developers in three field experiments [26]
3. 5 years average prior experience of METR trial developers on their own projects — the population that slowed down [5]
4. 28% vs 17% reduction in 'always-on' burnout-risk metric for developers without vs with Copilot [12]
5. 3 of 4 companies say the hardest part of generative-AI adoption is getting people to change how they work [8]
6. 40% maximum share of coders' work AI assistants may be able to take on (SaaS teams of 2,000–20,000 FTEs) [8]
7. 63% of developers say leaders do not understand their pain points (44% said leaders were aware in 2024) [9]
8. 62% view AI tools favourably (31.6% very, 30.4% somewhat); 53% of developers with 1–5 years’ experience vs 69% with 16+ years (n=14,428) [29]
9. 75.3% of developers go to a person when they don't trust an AI answer; 61.7% for ethical or security concerns (2025) [1]
10. 39% vs 14% expect AI to reduce headcount next year vs those who saw an actual AI-driven decline last year (all functions) [24]
11. 17.1% avoid AI at work or school for fear of losing skills or training their replacement — the top reason; 41.3% do not avoid it (n=13,857) [29]

Formulas and derived metrics

Derived quantities computed from the primary figures above. Use them to reason about agents rather than just quote them.

Frequently asked questions

Methodology

Primary sources are used wherever they exist: the organisation that ran the survey, benchmark or measurement. Secondary aggregators, vendor landing pages without a named methodology, and market-size forecasts without a disclosed model are excluded; the few secondary reports kept are labelled as such in the source list. Each figure is reproduced as stated by the source, with sample size and fielding dates where the source discloses them. Survey figures are self-reported by respondents and are not independently verified; benchmark figures depend on the harness and prompt used and should be compared only within a benchmark. The page is re-checked and updated as new editions of each report are published.

Sources